Communications and Public Relations
Two problems that cancelled each other out
A national public relations agency was burning tokens haphazardly and could not get its people to use AI.
Both, at the same time. Part of the firm was spending real money on AI without much to show for it. Most of the firm was not touching it at all. Either problem alone has an obvious answer, and together they cancel out, because the fix for one makes the other worse. Tighten the spending and you confirm what the holdouts already suspect. Push adoption and you multiply the waste.
What was at stake
Leadership named the asks without prompting. Increase practitioner output. Build consistency across every client. Build company memory that any deliverable or pitch could draw on.
Underneath sat something less comfortable. Agents were being stood up in a no-code builder and pointed at critical systems, with no service design behind them and nothing written down about what a machine was allowed to do.
And the practitioners were against it. Not indifferent. Against. In a craft business, where the product is judgment and voice and relationships, a workforce that believes the machine is coming for the thing they are good at will find ways to be right about that.
Leadership was not wrong
The usual version of this story has the floor disproving what the boardroom believed. That is not what happened.
The CEO and the managing partners had the diagnosis correct. They knew AI was being used badly and not used at all. They knew the mundane work was eating their people. What they did not have was a plan, or a way to run service design against their own operation to find out what the day-to-day work actually needed before anyone automated a piece of it.
The gap was method, not insight. That is the harder problem, because there is nothing to correct. There is only work to do.
A week on the floor
We spent a week in review, most of it interviewing practitioners directly.
There was no dramatic reveal. Leadership already knew media relations carried a lot of homework before anyone reached the work clients pay for. What the week established was the size of it.
Trace one pitch. A publicist starts with the news, manually, site by site, hunting the day's angle. One to two hours a day, every publicist. Then reporter research: who covers this, who has written something adjacent, who might take it. Only then does the publicist craft the pitch.
The craft is the third step. The first two run first, every day, by hand, before the work can start.
Govern first, then build
Most AI programs build something impressive and write the rules afterward. Here the order had to reverse, because the workforce was the real constraint.
So the first thing built was not an agent. It was a boundary.
AI output went into a separated directory that could not be used directly. Nothing produced there reached a client, a deliverable, or a pitch until a person had looked at it, exercised judgment, and graduated it to approved. The machine could work. The machine could not ship.
We did not invent that for this engagement. It is how we run AI inside every client we work with, which is why it could go first: it arrived already proven.
That one mechanism does three jobs. It answers the governance gap. It turns a promise into a structure, because we told the practitioners we would respect the craft and the people doing it, and the directory is what that sentence means in practice. And it settles what AI is for: a collaborator, not an autonomous doer.
The second decision was sequence. The first tool through the door was writing tooling we already had, proven elsewhere before this engagement. Not the most impressive thing available. The lowest-risk way to let a skeptical practitioner see what good looks like before anything autonomous entered the building.
Then a new business intake agent, taking the research burden off inbound leads. Then research, which is the answer to the first two steps of that pitch trace: the day's news gathered agentically and delivered as a report. Then a knowledge graph, the company memory leadership had asked for in the first conversation.
Training ran alongside all of it. Not prompt tricks. How to use AI effectively and responsibly, taught to the people with every reason to resist it.
What changed
The gathering stopped being human work. An agent reads across the sources and produces a report; the publicist reads the report and decides which angle is worth a pitch. Nobody hunts site by site any more, and the judgment about what is newsworthy never moved.
The machine took the gathering. It did not take the choosing. That is the same line the separated directory draws, and it is why the hours could collapse without anyone feeling replaced.
Adoption increased dramatically.
The more telling part is who is using it. This is the workforce that was against it. The people who believed the machine was coming for their craft are the ones running it, and they came around for earned reasons: they were trained properly, and the boundary held.
What they run without us
We stopped building.
Self-improvement and new agent building were taught to the firm's core enablement group, which we also trained. They build for themselves now. The capacity to make the next agent sits inside the agency rather than with us.
What continues is a standing monthly conversation with that group: questions, what to build next, and oversight of the governance that made the rest possible.
The engagement did not end. It turned into advice, which is what capability over dependency looks like when it works.